Most API pricing pages quote three numbers per million tokens, and it's easy to read them as "this one is $1 more expensive". The input prices of Gemini 4 Neon and Claude Opus 5.5 differ by exactly $1 per million tokens. Over a year of realistic traffic, that $1 becomes $240.
Here is the arithmetic, so you can redo it with your own token mix.
The only formula you need
monthly = input_tokens × input_price + output_tokens × output_price
Both prices are per 1M tokens. Cached input is a separate rate, billed only on the portion that hits the cache. That's the whole model — no tiers, no minimums in the published tables I used.
A fixed workload, so the numbers are comparable
Pick a workload and hold it constant across models: 10M input tokens and 2M output tokens per month, no cache hits. 12M tokens total.
| Model | Input /1M | Cached input /1M | Output /1M | Monthly | Yearly |
|---|---|---|---|---|---|
| Gemini 4 Neon | $3.00 | $0.15 | $15.00 | $60.00 | $720.00 |
| Claude Opus 5.5 | $4.00 | not published | $20.00 | $80.00 | $960.00 |
| Claude Opus 5.5 (fast) | $8.00 | not published | $40.00 | $160.00 | $1,920.00 |
| GPT‑6.2 Sol | $2.00 | $0.10 | $10.00 | $40.00 | $480.00 |
Every rate above comes from the vendor's own pricing page. Neon and both Opus rows were read on 2026‑11‑18; Sol was read on 2026‑11‑20. "not published" means the vendor's page does not list a cached-input rate for that model — I'm not going to estimate one.
Finding 1: output tokens are a sixth of the volume and half the bill
On Neon: 10M input × $3.00 = $30, and 2M output × $15.00 = $30. Exactly half the bill comes from 16.7% of the tokens.
That isn't a Neon quirk. Every row in the table above prices output at exactly 5× its input rate, and in every row the output line lands at 50% of the total. The ratio of output to input is what decides your bill, and it's usually the number nobody measures. If your app is chat-shaped, output share is often higher than 1:5.
Finding 2: cached input is where the honest comparison breaks
Neon lists a cached-input rate of $0.15 per 1M, which is 5% of its $3.00 input rate. Opus 5.5 doesn't publish one yet.
This matters the moment your prompts carry a long shared prefix — a system prompt, a tool schema, a document you keep resending. If 80% of your input hits cache, the input side of the Neon row falls from $30 to roughly $7.20 and the model gets a lot cheaper. For Opus 5.5 I can't run that variant at all, because the rate isn't public. Two models with a published cached rate can be compared on cache-heavy traffic; one with and one without cannot. Don't let a calculator quietly assume zero cache hits for one and the published rate for the other.
Finding 3: the "fast" tier doubles everything
Opus 5.5 (fast) is $8.00 in / $40.00 out — cleanly 2× the standard tier. Same 12M-token workload: $160/month, $1,920/year, which is $100/month more than Neon and 167% higher. If you're testing a latency-sensitive feature, that multiplier belongs in the design doc, not in the invoice surprise.
How to redo this for your own numbers
- Pull last month's actual input and output token counts from your provider's usage dashboard. Don't guess the split.
- Multiply each by the per-1M rate from the vendor's pricing page, and write down the date you read it.
- If you use caching, split input into cached and uncached and price them separately.
- Re-run when a vendor changes a rate. Prices move; the formula doesn't.
Sources and caveats
- Gemini 4 Neon rates: Google's pricing page, read 2026‑11‑18.
- Claude Opus 5.5 and 5.5 (fast) rates: Anthropic's pricing page, read 2026‑11‑18.
- GPT‑6.2 Sol rates: OpenAI's pricing page, read 2026‑11‑20.
- All yearly figures are monthly × 12. No volume discounts, committed-use pricing, or free tiers are applied, because none of the pages I read listed them for these models.
Vendor prices change without much notice, so check the source page before you put a number in a budget. I keep the full matrix — every rate, with the date it was read — at https://modelcost.org/, and there's a calculator there if you'd rather move sliders than multiply. If you find a rate that's wrong or out of date, write to hello@modelcost.org and I'll fix it.
Not affiliated with Google or Anthropic. Model names are used descriptively only.
Disclosure: this draft was written with AI assistance. Every price and calculation above was checked by hand against the vendor pricing pages listed, on the dates shown.
